Hybrid Training for 5G Linearization Systems Based on Machine Learning

Radio over Fiber systems blend the efficiency of fiber optic transmission with the flexibility of wireless communications but face significant challenges due to inherent nonlinearities in optical signal transmission. This paper proposes a hybrid training technique using simulated and real data to train machine learning models deployed for linearization, aiming to enhance performance and reliability. The proposal involves pre-training the model with a large synthetic dataset and fine-tuning it with real data collected from experiments. This method leverages synthetic data to establish initial weights, refined using real-world data to capture practical complexities. Results show that the hybrid training technique significantly outperforms models trained solely on synthetic data, with nearly a twofold improvement in performance as evidenced by lower Mean Squared Error (MSE) and Mean Absolute Error (MAE) values. Compared to models trained exclusively on real data, the hybrid method achieves comparable performance, proving effective in scenarios where real data is scarce or hard to obtain.